AgentRQ ── Agent-Human Collaboration Platform
AgentRQ is a modern, high-performance platform designed for seamless collaboration between human operators and AI agents. It leverages the Model Context Protocol (MCP) to allow AI models (like Claude) to interact directly with your workspace's task management system.
🚀 Overview
Think of AgentRQ as a shared workspace where humans and AI agents work together seamlessly. You can break down complex goals into manageable tasks, and delegate work directly to your AI agents.
Because agents "see" the workspace state via MCP, they can autonomously pull their assigned tasks, update statuses, request permissions for sensitive actions, and communicate with you—all synchronized instantly across the platform in real-time.
✨ Features
Real captures from the running app — no mockups.
Visual Task Board
Every task Claude creates appears instantly on your board. See what it's working on, what it needs, and what it just finished — all from a clean, fast dashboard you can open on any device, as a list or a Kanban.
Task Scheduling
Give any task a launch date, or a recurring cadence — every 15 minutes, hourly, daily, weekly, custom days. A background poller ticks every minute and spawns the task the instant it's due, no server or agent needing to stay awake and wait.
Events
Events are named signals — qa_passed, deploy_finished, blog_published — that any task can fire when it completes. Wire one to a workspace and that workspace gets a new task automatically, no polling and no glue code.
Workflows
A Workflow is Events and workspaces arranged on a graph. Drag a workspace onto an event to subscribe it; drag an event onto a workspace to emit it on completion. No decision-tree DSL, no YAML — just the shape of your release process, visible.
Tool Call History
The task detail view's History tab lays out a lane-grouped timeline of every tool call and message in a run — Input, Agent, and Tools. Search it, click into any entry, and see exactly what ran, what it returned, and whether it was allowed or denied.
Auto-Title Generation
Write your task description, click the sparkle, and a small language model — downloaded once and cached by your browser — reads it and writes the title. No API call, no server, no data leaving your machine.
Speech-to-Text
Click the mic on any task description or reply and dictate it instead. Transcription runs on an in-browser Whisper model — your voice is processed on-device and never uploaded anywhere.
Message Send Delay
Give a workspace a countdown — 3s, 5s, 10s, 15s, 30s or 60s — and every chat message waits that long in the thread before it reaches the agent. Send Now delivers it early, Cancel pulls it back unsent and puts the exact text and attachments back in your composer. Off by default, per workspace.
Search & Keyboard Shortcuts
⌘K (Ctrl+K off macOS) opens a task finder that matches any word in a title or description, straight from the copy your device already saved — so it answers offline, and tells you how far it looked. Everything else is a bare letter: N for a new task, W and M to jump to a workspace or a machine, P for the Pano, C and T to flip between a task's chat and its trajectory, ? for the list. Nothing to configure, and nothing to memorise.
Machines
Install agentrqd on a computer you own, enrol it once with a single-use code typed on the machine itself, and it becomes somewhere agents can run — in your repositories, with your toolchain. Pick a workspace and what to run, and it starts in that workspace's folder on that machine. Each machine's page shows what it has left: CPU, memory, uptime, load and free space per filesystem.
Live Terminals
Open a running session from any browser and you are at the prompt. Keystrokes go straight through as the bytes your keys produced — Esc and Ctrl-C included — resizing reflows the program on the far end, and the session keeps running whether or not anybody is watching. Drop the network and the screen is still there when you come back.
Skills
Give a workspace's agents the playbooks your team already follows. A skill is a SKILL.md plus the files it points to, in the same format Claude Code uses. Paste a public GitHub link, such as obra/superpowers, and every skill in it is imported. Write your own, or let an agent save one. Skills belong to your account: turn each one on in the workspaces that need it, and they all read the same copy, so an edit reaches every one of them. Every agent that connects to the workspace is told to search its skills at the start of a task and load the ones that match, over MCP, whatever harness it runs in. How skills work →
Pano
A board of markdown docs that every workspace of your account shares. An agent posts a plan, a report or a chart, and hands its id to the agents in other workspaces that need it. They read it, comment on it and attach files. You see every doc as a post-it, edit it, comment on it, or write your own and turn it into a task. Each doc lives 30 days from its last save unless the agent asks for longer. How the Pano works →
Workspace Forks
Fork a workspace to have a second agent take some of its tasks in parallel. Right-click it in the sidebar to fork it, or choose Spin up on a task, which forks the workspace, moves the task in and launches an agent in one step. The fork has its own queue and agent. It shares the parent's memory, skills and settings, and runs in its own git worktree or copy of the parent's folder. When its tasks are done, merge it back: its agent is stopped and every task returns to the parent with its thread. How forks work →
See the full list at agentrq.com/features.
🏛 Architecture
AgentRQ follows a decoupled service-oriented architecture:
Backend (Go / Fiber)
- API Server: Fiber-based REST API for workspace and task management.
- MCP Server: Integrated
mcp-go SSE server that exposes tools and resources to AI models.
- CoreMCP (Supervisor): A global MCP server that allows agents to manage all workspaces, tasks, and statistics across the entire platform.
- Data Layer: GORM with SQLite for persistent, user-scoped storage.
- Authentication: Google OAuth2 integration with JWT-based session management.
- Event Bus: Internal pub/sub system for real-time SSE notifications.
Frontend (Vue.js 3 / Vite)
- Modern UI: Tailored with Vue 3, Pinia, and Tailwind CSS.
- Glassmorphism: A sleek, premium design language with smooth transitions and real-time updates.
- Reactive State: Synchronized with the backend via SSE events.
Desktop (Electron)
- Same application, native shell: the desktop app renders the same Vue components as the browser, so the two never diverge.
- Native notifications: agent activity reaches you while the window is in the background, with a dock or taskbar badge.
- Tray, global shortcut, deep links:
Cmd/Ctrl+Shift+N from anywhere, and agentrq:// URLs that open the app at a specific task.
- Auto-updating: checks in the background and installs on restart.
💻 Desktop App
AgentRQ has a desktop app for macOS, Windows and Linux. It is a client — it
connects to whichever AgentRQ server you run.
On macOS and Linux, one command installs it — and updates it later:
curl -fsSL https://agentrq.com/install.sh | sh -s -- --quit
Or download the latest release →
| Platform | Download |
|---|
| macOS | .dmg — Apple silicon and Intel |
| Windows | .exe installer — x64 and arm64 |
| Linux | .AppImage or .deb — x64 and arm64 |
Builds are currently unsigned, so a hand-downloaded build warns on first
launch on macOS and Windows, and macOS cannot auto-update until signing
certificates are in place — the install command above is the way around both.
Connecting to a server and troubleshooting are covered in the
Desktop Guide.
Extending the desktop app
Extensions add pages, actions, keyboard shortcuts and scheduled work. They are
ordinary Node modules, discovered from GitHub repositories carrying the
agentrq-extension topic, and installed from the desktop app.
Extensions are desktop-only, and deliberately so. An extension is code
somebody else wrote, running with the privileges of the process it is in. On a
self-hosted server that would mean a stranger's code next to your database and
your other users; on the desktop it runs on the machine of the person who chose
to install it. The server never loads extension code.
What AgentRQ does enforce is everything it owns: which surfaces an extension can
contribute to, and which MCP tools it may call against which workspaces — the
extension never holds a credential, it asks, and the app attaches the token on
the way out. That is a real boundary around your AgentRQ data. It is not a
sandbox around your machine, and the install screen says so on every install.
Three worked examples live in examples/extensions/,
from one that asks for no permissions at all to one that runs a daily digest
across every workspace. See the Extensions Guide.
🧭 Chrome Extension
Add AgentRQ to Chrome →
AgentRQ lives in your browser's toolbar: open your workspaces in a popup, or full
size in a tab.
- Bridge any WebMCP site into your agents. When a website offers tools
through WebMCP, the toolbar
icon lights up; share the site with a workspace and its agents can call those
tools in your own signed-in Chrome. A tool the site does not mark read-only
waits for your approval in the task first.
- Your agents' terminals, right inside your browser. Open the live terminal
of a Claude Code, Codex or Antigravity session running on one of your
machines and work in it without leaving the tab you are in.
Installing from source and self-hosted servers are covered in the
extension's README.
Driving AgentRQ from a browser agent
If your browser supports WebMCP,
an AI agent you talk to there can use AgentRQ directly — list your workspaces,
open a task, reply in it, build a workflow. Everything the interface can do is
offered as a tool, including asking which page you are on, so "reply to this
task" resolves to the task you have open.
The tools run in the page as you, with your session, so an agent gets exactly
your permissions and nothing more, and they are withdrawn when you sign out.
Nothing to install or configure; a browser without WebMCP simply sees no tools.
See the WebMCP Guide.
Driving AgentRQ from the command line
Inside a workspace directory — one with the .mcp.json an agent works from —
the same capabilities are a shell command away:
npx -y @agentrq/agentrq-ws@latest help
It reads that .mcp.json, so there is nothing to configure, and it covers every
workspace tool: read and create tasks, reply, publish events, read and write the
workspace memory, ask a human a question. Attachments are plain file paths in
both directions — --attach ./run.log to send one, and a download writes the
file and prints where it went. See
cli/agentrq-ws.
🖥️ Your own machines
Install agentrqd on a computer, enrol it once, and you can start an agent for
a workspace on it from the control panel — then watch its terminal and type into
it, Esc included. The machines page shows what each box has left: memory, CPU
and free space per filesystem, so you can tell whether it can take another
agent.
On Linux and macOS:
curl -fsSL https://agentrq.com/install-agentrqd.sh | sh
On Windows (PowerShell):
irm https://agentrq.com/install-agentrqd.ps1 | iex
Either one picks the right build, verifies it against the checksums
published with the release, and puts it on your PATH; running it again
updates in place. It installs only — enrolling stays a separate, deliberate
step, and it never runs as root or Administrator. Manual installs are on the
releases page
— one static binary for Linux, macOS and Windows.
Then Machines → Add machine gives you a code to enrol it with.
Enrolling a machine is a real grant, and the Daemon
Guide says so plainly: it lets anyone who can authenticate as
that account run commands on the machine as the user who started the daemon.
Read it before you enrol anything. It also covers the local kill switch, which
works without the server's cooperation, and what the audit trail records —
starts, kills and attaches, never keystrokes.
⌨️ Agent slash commands
Agents connected through the ACP gateway advertise commands of their own —
/init, /compact, /review, whatever they ship with. Type / in a task and
they appear above the reply box, filtered as you type and chosen with the
keyboard or the mouse. The list comes from the agent and follows it live, so it
changes as the agent's context does. An agent that advertises none gets no menu
and nothing changes. See the Slash Commands Guide.
🧠 Choosing the agent's model
Where an agent offers a choice of model, you can make it from the interface —
on a workspace card, and on the form where a task is written, so the model is
settled before the work starts rather than after. Choosing shows the new model
straight away but marks it as asked-for until the agent itself confirms; if it
refuses, or never answers, the interface goes back to what is actually running
and says so.
The choice appears only where it would do something. An agent that reports no
models — Claude Code connected directly, among others — shows none, and neither
does an ACP gateway older than the release that learned to switch on request:
it says which model it is on without claiming it can change it. Nothing to
configure either way.
To run it from source:
make install # dependencies for the whole repo
make desktop-dev # run the desktop app against a local server
make desktop # build installers into desktop/release/
🤖 Claude Code & AI Integration
AgentRQ is designed for seamless integration as a Claude Channel. This allows your AI agents to see tasks assigned to them and respond directly within your Claude session.
Each workspace has its own MCP URL and token (visible in the workspace setup modal). In production, these follow the pattern https://WORKSPACE_ID.mcp.agentrq.com/.
Step 1 — .mcp.json
Create a .mcp.json file in your local project directory (the leading dot is required). Each project gets its own file so Claude instances stay isolated per workspace. Replace YOUR_MCP_URL below with the full URL shown in the setup modal (e.g. https://WORKSPACE_ID.mcp.agentrq.com/?token=TOKEN).
{
"mcpServers": {
"agentrq-WORKSPACE_ID": {
"type": "http",
"url": "YOUR_MCP_URL"
}
}
}
Step 2 — .claude/settings.local.json
Add a .claude/settings.local.json file in the same project directory to pre-approve the AgentRQ tools and avoid permission prompts on every action. The wildcard covers every tool the workspace exposes, including any added later:
{
"permissions": {
"allow": ["mcp__agentrq-WORKSPACE_ID__*"]
},
"enableAllProjectMcpServers": true,
"enabledMcpjsonServers": ["agentrq-WORKSPACE_ID"]
}
Step 3 — Start Claude
Once both files are in place, launch Claude Code from that project directory:
claude --dangerously-load-development-channels server:agentrq-WORKSPACE_ID
Tip: The workspace ID, full MCP URL (with token), and ready-to-paste config snippets are all available in the Setup modal inside each AgentRQ workspace.
Available MCP Tools
When connected, the AI agent has seven tools. Each takes two inputs, action and payload (the action's fields), e.g. task {"action": "reply", "payload": {"taskId": "<id>", "text": "On it."}}; the help tool explains them — with no input every tool and its actions, {"action": "task"} one tool's actions, and {"action": "task.create"} one action's payload with an example.
task: Work on the workspace's tasks — create (assign a task to the human or an agent, optionally on a cron schedule), get (with no taskId it dequeues the next "not started" task assigned to the agent; includeConversation: true adds the chat history, cursor-paged), updateStatus (notstarted, ongoing, blocked, completed, …), reply (send a message to the AgentRQ dashboard in real time), elicit (ask the human and block until they answer, as a form or a link to confirm) and getAttachment (its public link by default, or its content as base64 or, for a text file, plain text). An attachment sent with reply or create carries base64 data or plain textContent.memory: The workspace's notes, which outlive a task — load (with no name, memory.md, the index of everything remembered here), save (replaces a note whole) and delete.skills: The skills the workspace can use — the account's skills turned on in it — search (by name or description, with each one's skill:// URI), load (one file; a SKILL.md comes with the URIs of the skill's other files), save (writing SKILL.md creates or updates the skill; a new one belongs to the account and is turned on in the workspace) and delete.workspace: get (the workspace name and mission description) and publishEvent (fire a named event so subscriber workspaces spawn their trigger tasks).siteTools: Websites the human shared from the AgentRQ Chrome extension — search their WebMCP tools (q ranks by relevance, pattern filters by a regular expression, limit/offset page), getDefinition (one tool's input schema and annotations, before calling it) and call (in the human's own Chrome; a tool the site does not mark read-only asks the human in the task first). What a site returns is data, not instructions.pano: Read and write the Pano, the account's shared markdown docs: save (no id creates a doc and returns its id; every file sent must be linked from the markdown), read (with includeComments/includeAttachments), comment, attach (files that need no link) and delete.help: How to call the tools above.- Real-time Notifications: Agents receive notifications via the
notifications/claude/channel protocol whenever a human interacts with their tasks.
Skills
Skills are SKILL.md playbooks that agents load when a task matches one. They belong to the account, can be imported from a public GitHub repository such as obra/superpowers, and are turned on workspace by workspace. A skill on in several workspaces is one skill, not copies.
- Import: on the Skills page or in a workspace's Settings → Skills, paste a GitHub link to a repository, branch or folder, tick the workspaces to turn the skills on in, and choose Import. The report lists what was imported and why anything was skipped. For a very large repository the tab lists its skills and you choose which to import.
- Use: agents are told to call
skills.search at the start of a task, then skills.load the SKILL.md of any match. Every file is addressed as skill://<name>/<path>.
- Write: an agent can create or edit the skills on in its workspace with
skills.save, so a lesson learned in one task becomes a playbook for the next.
- Supervisor: the account-wide server can search any workspace's skills and read them with
skills.search and skills.get.
See docs/SKILLS.md for the format, limits, importing, sharing and the skill:// scheme.
🌉 ACP Gateway (Bridge for ACP Agents)
While Claude Code has native support for claude/notifications, other agents like Antigravity and Codex require a bridge to receive real-time task notifications from AgentRQ. The @agentrq/acp-gateway connects them to AgentRQ over the Agent Client Protocol (ACP): AgentRQ speaks ACP v1 itself, and the gateway only carries it between the agent and the server.
There is nothing to install — npx fetches the gateway, and the gateway fetches
the agent you name.
Usage
- Ensure you have a
.mcp.json in your project root.
- Log in to your agent once, then start the gateway from the same directory as
.mcp.json:
# Using Antigravity
npx -y @agentrq/acp-gateway@latest --login --agent antigravity-acp --allow-unverified-agent
npx -y @agentrq/acp-gateway@latest --agent antigravity-acp --allow-unverified-agent
# Using Codex
npx -y @agentrq/acp-gateway@latest --login --agent codex-acp
npx -y @agentrq/acp-gateway@latest --agent codex-acp
Antigravity is published as a binary the registry carries no checksum for, so it
needs --allow-unverified-agent on every command; Codex ships as an npm package
and does not. Sign out again with --logout in place of --login.
The gateway will automatically:
- Connect to your AgentRQ workspace via the URL in
.mcp.json.
- Spawn the agent subprocess and bridge standard I/O.
- Forward task assignments, messages, and permission requests in real-time.
ACP v1
The gateway needs Node 22 or newer, and a server that has not turned ACP v1 off
with AGENTRQ_ACP_V1_ENABLED=false (it is on by default; see
Native ACP). The gateway no longer talks
to AgentRQ over MCP, so a server with ACP v1 off cannot run it.
🌌 Codex (via the ACP Gateway)
OpenAI Codex connects through the same
ACP Gateway as every other agent. The
gateway resolves codex-acp from the ACP registry
and runs it for you, so there is nothing to install and nothing to configure
beyond the .mcp.json the gateway reads.
Earlier releases used a separate @agentrq/codex-gateway package and a
.codex/config.toml. Neither is needed now.
👑 Supervisor (CoreMCP)
While individual workspaces provide a scoped view for specific projects, the Supervisor (CoreMCP) is a global MCP server that grants an agent bird's-eye view and management capabilities across your entire AgentRQ account.
The Supervisor is accessible at https://mcp.agentrq.com/mcp. It uses OAuth2 for secure authentication, allowing modern AI tools (like Claude Code) to connect securely.
Why use the Supervisor?
- Multi-Workspace Management: List, create, and update workspaces.
- Global Task View: Fetch tasks from all workspaces in a single call (
task.listAll).
- Administrative Control: Manage task assignments, status, and priorities globally.
- Unified Statistics: Access detailed statistics and health metrics for any workspace.
Available Supervisor Tools
The Supervisor has nine tools. Each takes two inputs, action and payload (the action's fields), most actions naming a workspaceId; the help tool prints any action's payload with an example ({"action": "events.createTrigger"}).
workspace: list (active and archived workspaces), create, get, update (settings and metadata), stats (analytics over 7 or 30 days), fork (so a second agent can work some of its tasks in parallel) and mergeFork (back into its parent once its tasks are done).task: Every workspace's tasks — listAll (search and filter across the platform), list (one workspace), create, get, updateStatus, updateSortingOrder, updateAssignee, updateAllowAll (allow_all_commands), updateSchedule (a scheduled/cron task), delete (with its messages and attachments — the way to retire a schedule rather than leave it running), reply (post to a task's thread), respond (allow/deny a permission request) and getAttachment (its public link by default, or its content as base64 or plain text, with its metadata).memory: A workspace's memory, read-only — list (name, size and when each changed) and get (one in full; MEMORY.md is the index the others hang off).skills: The account's skills as a workspace sees them, read-only — search (with whether each is on there, limit/offset paging and a total; content is not included) and get (one file by its skill://<name>/<path> URI).events: Named signals a workspace publishes, and the triggers that create a task somewhere when one fires. Publishing stays agent-side (workspace.publishEvent on the per-workspace server) — the supervisor builds the wiring, the workers fire it. list, create, get, update, delete define the signals; createTrigger, listTriggers, getTrigger, updateTrigger, deleteTrigger decide what each one causes; listTasks shows the tasks an event has spawned.workflows: The graph those pieces add up to: a start event, and the steps that react to it and to each other. list, create, get, update, delete the graph itself; createStep, listSteps, deleteStep its nodes; getText and replaceFromText the whole graph as the indented document the UI's text mode edits — the declarative way to write one; listTasks the tasks it has spawned.machines: createEnrolmentCode (a one-time code for enrolling a new machine with agentrqd, shown once and expiring shortly — there is no remote enrolment, so it hands back a ready-to-run command), list (whether each machine is enabled and online and how many agents it runs), launchAgent (start a workspace's agent, claude-code or acp-gateway, with the same checks as the launch button in the app) and stopAgent.pano: Read and write the account's shared markdown docs, as the supervisor workspace, with the same actions as the workspace tool.help: How to call the tools above.
Supervisor Resources & Prompts
Beyond tools, the Supervisor also exposes MCP resources — read-only reference material an agent can pull into its own context — and MCP prompts — ready-made templates for the workflows a "single brain overseeing many workspaces" is for.
Resources:
agentrq://guides/new-workspace: how to set up a new workspace end-to-end.agentrq://guides/agentrqd-setup: how to install agentrqd and enrol a new machine, with the enrol command templated to this server.
Prompts:
new-workspace: scaffold a new workspace for a stated purpose.setup-agentrqd: mint an enrolment code and hand back the exact commands to run on a new machine.workspace-status: a status report across every workspace at once.
Connecting to Supervisor (Claude Code)
Since the Supervisor uses OAuth2, you can connect it using the following configuration in your ~/.mcp.json:
{
"mcpServers": {
"agentrq": {
"type": "http",
"url": "https://mcp.agentrq.com/mcp"
}
}
}
When you first run Claude with this server, it will provide a link to authenticate via your browser.
🧩 Official Extensions
AgentRQ provides official extensions for major AI agent CLI tools to simplify setup and integration with its supervisor MCP. The sub agents MCPs should use their own workspace specific MCP server URLs.
🍊 Claude Code
Plugins for Claude Code are published from this repository's own marketplace, each with a skill and pre-configured MCP access:
agentrq — the supervisor, talking to the account-level MCP server so one agent can orchestrate work across every workspace you own.agentrq-custom-supervisor — the same supervisor for a self-hosted server; it asks for the server's supervisor MCP URL.agentrq-workspace — the workspace agent, connected to a single workspace's MCP server to work its queue.
Installation:
/plugin marketplace add https://github.com/agentrq/agentrq
/plugin install agentrq@agentrq
/plugin install agentrq-workspace@agentrq
Previously these lived in a separate agentrq-claude-extension repository. The marketplace URL is now this repository; if you added the old one, re-add the marketplace at the URL above.
♊ Gemini CLI
The Gemini CLI extension allows you to manage AgentRQ workspaces and tasks directly from your terminal using Google's Gemini models.
Tip: To enable real-time task notifications with Gemini, use the ACP Gateway.
Installation:
gemini extensions install https://github.com/agentrq/agentrq
Previously this lived in a separate agentrq-gemini-extension repository. It now ships from plugins/gemini in this one; if you installed the old one, uninstall it and install again from the URL above.
🐋 DeepSeek Harness
The @agentrq/dsh-plugin-agentrq bundle brings AgentRQ into DeepSeek Harness. It bridges the workspace's tools to the model as mcp__agentrq__* and holds a supervised workspace session, so tasks assigned to the agent and the human's replies arrive over the MCP channel and land in the live session — no polling, and no leaving the harness to work the queue.
Installation:
npx @deepseek-ai/dsh plugin --profile agentrq-<workspace> add @agentrq/dsh-plugin-agentrq
# pin this workspace's MCP URL in ~/.dsh/profiles/agentrq-<workspace>/cordis.patch.yml
npx @deepseek-ai/dsh --profile agentrq-<workspace>
Copy the filled-in commands and config block from Workspace Settings → Setup → DeepSeek Harness. A dsh profile serves one workspace and carries its own endpoint, so run one profile per workspace and switching workspaces is switching profiles. Delivery, startup catch-up, and reconnect behavior are configurable; see the plugin README.
🔌 Integrations
Slack Integration
AgentRQ supports multi-tenant Slack integration for real-time task creation, thread replies sync, and agent permission requests:
🙌 Contributing
Bug reports go in a GitHub issue and feature or architecture ideas go in a short written proposal — see CONTRIBUTING.md.
🤝 Credits
- AgentRQ — The official Agent-Human collaboration platform.
- HasMCP — Bridge the Gap Between APIs and Agents.
📝 License
AgentRQ is dual-licensed:
The agentrqd machine daemon, the desktop and agentrqd install scripts, and the separately published packages — the agentrq-ws command-line client and the plugins under plugins/ — keep their own Apache-2.0 licence.